Abstract
The Adaptive MaxWeight policy achieves optimal throughput for switches with nonzero reconfiguration delay and has been shown to have good delay performance in simulation. In this paper, we analyze the queue length behavior of a switch with nonzero reconfiguration delay operating under the Adaptive MaxWeight. We first show that the Adaptive MaxWeight policy exhibits a weak state space collapse behavior in steady state, which can be viewed as an inheritance of a similar property of the MaxWeight policy in a switch with zero reconfiguration delay. The weak state space collapse result is then utilized to obtain an asymptotically tight bound on an expression involving the steady-state queue length and the probability of reconfiguration for the Adaptive MaxWeight policy in the heavy traffic regime. We then derive the relation between the expected schedule duration and the steady-state queue length through Lyapunov drift analysis and characterize bounds for the expected steady-state queue length. While the resulting queue length bounds are not asymptotically tight, they suggest an approximate queue length scaling behavior, which approaches the optimal scaling with respect to the traffic load and the reconfiguration delay when the hysteresis function of the Adaptive MaxWeight policy approaches a linear function.
Similar content being viewed by others
Notes
From the definition of the MaxWeight schedule, we know that for any maximal schedule \(\mathbf {s}'\) that is not a maximal schedule, there always exists a maximal and MaxWeight schedule \(\mathbf {s}^*\) which covers all the queues served by \(\mathbf {s}'\), hence we can assume that the MaxWeight policy always generates a maximal schedule without loss of generality. It then follows that the Adaptive MaxWeight always generates a maximal schedule since its schedule is generated by the MaxWeight policy, either at the current or a previous time slot.
References
Bertsimas, D., Gamarnik, D., Tsitsiklis, J.N.: Performance of multiclass Markovian queueing networks via piecewise linear Lyapunov functions. Ann. Appl. Probab. 11, 1384–1428 (2001)
Bramson, M.: State space collapse with application to heavy traffic limits for multiclass queueing networks. Queueing Syst. 30(1), 89–140 (1998). https://doi.org/10.1023/A:1019160803783
Çelik, G., Borst, S., Whiting, P., Modiano, E.: Variable frame based max-weight algorithms for networks with switchover delay. In: 2011 IEEE International Symposium on Information Theory Proceedings, pp. 2537–2541 (2011)
Celik, G., Borst, S.C., Whiting, P.A., Modiano, E.: Dynamic scheduling with reconfiguration delays. Queueing Syst. Theory Appl. 83(1–2), 87–129 (2016). https://doi.org/10.1007/s11134-016-9471-4
Chan, C.W., Armony, M., Bambos, N.: Maximum weight matching with hysteresis in overloaded queues with setups. Queueing Syst. Theory Appl. 82(3–4), 315–351 (2016). https://doi.org/10.1007/s11134-015-9468-4
De Dobbelaere, P., Falta, K., Gloeckner, S., Patra, S.: Digital mems for optical switching. IEEE Commun. Mag. 40(3), 88–95 (2002). https://doi.org/10.1109/35.989763
Eryilmaz, A., Srikant, R.: Asymptotically tight steady-state queue length bounds implied by drift conditions. Queueing Syst. Theory Appl. 72(3–4), 311–359 (2012). https://doi.org/10.1007/s11134-012-9305-y
Kang, W.N., Kelly, F.P., Lee, N.H., Williams, R.J.: State space collapse and diffusion approximation for a network operating under a fair bandwidth sharing policy. Ann. Appl. Probab. 19(5), 1719–1780 (2009)
Keyworth, B.: Roadm subsystems and technologies. In: Optical Fiber Communication Conference, 2005. Technical Digest. OFC/NFOEC, vol. 3, pp. 3–4 (2005). https://doi.org/10.1109/OFC.2005.192706
Li, X., Hamdi, M.: On scheduling optical packet switches with reconfiguration delay. IEEE J. Sel. Areas Commun. 21(7), 1156–1164 (2003). https://doi.org/10.1109/JSAC.2003.815843
Li, Y., Panwar, S., Chao, H.: Frame-based matching algorithms for optical switches. In: 2003, HPSR. Workshop on High Performance Switching and Routing, pp. 97–102 (2003). https://doi.org/10.1109/HPSR.2003.1226687
Liu, H., Lu, F., Forencich, A., Kapoor, R., Tewari, M., Voelker, G.M., Papen, G., Snoeren, A.C., Porter, G.: Circuit switching under the radar with reactor. In: Proceedings of the 11th USENIX Conference on Networked Systems Design and Implementation, NSDI’14, pp. 1–15. USENIX Association, Berkeley, CA, USA (2014). http://dl.acm.org/citation.cfm?id=2616448.2616450
Liu, H., Mukerjee, M.K., Li, C., Feltman, N., Papen, G., Savage, S., Seshan, S., Voelker, G.M., Andersen, D.G., Kaminsky, M., Porter, G., Snoeren, A.C.: Scheduling techniques for hybrid circuit/packet networks. In: ACM CoNEXT, ACM CoNEXT’15. ACM, New York (2015)
Maguluri, S.T., Burle, S.K., Srikant, R.: Optimal heavy-traffic queue length scaling in an incompletely saturated switch. SIGMETRICS Perform. Eval. Rev. 44(1), 13–24 (2016). https://doi.org/10.1145/2964791.2901466
Maguluri, S.T., Srikant, R.: Heavy traffic queue length behavior in a switch under the maxweight algorithm. Stoch. Syst. 6(1), 211–250 (2016). https://doi.org/10.1214/15-SSY193
McKeown, N., Anantharam, V., Walrand, J.: Achieving 100% throughput in an input-queued switch. In: INFOCOM’96. Fifteenth Annual Joint Conference of the IEEE Computer Societies. Networking the Next Generation. Proceedings IEEE, vol. 1, pp. 296–302 (1996). https://doi.org/10.1109/INFCOM.1996.497906
Tassiulas, L., Ephremides, A.: Stability properties of constrained queueing systems and scheduling policies for maximum throughput in multihop radio networks. IEEE Trans. Autom. Control 37(12), 1936–1948 (1992). https://doi.org/10.1109/9.182479
Venkatakrishnan, S.B., Alizadeh, M., Viswanath, P.: Costly circuits, submodular schedules and approximate carathéodory theorems. In: Proceedings of the 13th ACM SIGMETRICS/PERFORMANCE Joint International Conference on Measurement and Modeling of Computer Systems, SIGMETRICS’16. ACM, New York (2016)
Wang, C., Javidi, T.: Adaptive policies for scheduling with reconfiguration delay: an end-to-end solution for all-optical data centers. IEEE/ACM Trans. Netw. 25(3), 1555–1568 (2017). https://doi.org/10.1109/TNET.2016.2644617
Wang, C.H., Javidi, T., Porter, G.: End-to-end scheduling for all-optical data centers. In: INFOCOM, 2015 Proceedings IEEE (2015)
Wang, W., Maguluri, S.T., Srikant, R., Ying, L.: Heavy-traffic delay insensitivity in connection-level models of data transfer with proportionally fair bandwidth sharing. SIGMETRICS Perform. Eval. Rev. 45(3), 232–245 (2018). https://doi.org/10.1145/3199524.3199565
Acknowledgements
This material is based upon work supported by the National Science Foundation under Grants CNS-1329819 and CCF-1850439.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendices
Proof of Proposition 1
Proof
For ease of notation, we drop the superscript \((\epsilon )\) in the following derivation. For each state \(\mathbf {X} = (\mathbf {q}, \mathbf {s}, r)\), we define the Lyapunov function \(Z(\mathbf {X}) = \max \{ \Vert \mathbf {q}_{\perp }\Vert - \theta \Vert \mathbf {q}_{\parallel } \Vert , 0 \}\). We then apply Lemma 1 with the Lyapunov function Z to obtain the result. Note that the selection of the Lyapunov function is such that Z is a nonnegative function. Since \(\Vert \mathbf {q}_{\perp }\Vert - \theta \Vert \mathbf {q}_{\parallel } \Vert \le Z(\mathbf {X})\) for any state \(\mathbf {X} = (\mathbf {q}, \mathbf {s}, r)\), the statement of Proposition 1 follows from a bound on \({\mathbb {E}}[Z(\bar{\mathbf {X}})]\).
We first verify Condition C.2 for \(Z(\mathbf {X})\). Since \(|\max \{a, 0\} - \max \{b, 0\}| \le |a - b|\) for any \(a, b \in {\mathbb {R}}\), we have
Here, we use the fact that \(\mathbf {q}_{\perp }\) is a projection onto \(\mathcal {K}^{\circ } = \{\mathbf {x} \in {\mathbb {R}}^{n^2}: \langle \mathbf {x}, \mathbf {y} \rangle \le 0, \forall \mathbf {y} \in \mathcal {K} \}\), the polar cone of \(\mathcal {K}\). Since the projection onto a cone is nonexpansive, we have \(\Vert \mathbf {x}_{\perp } - \mathbf {y}_{\perp } \Vert \le \Vert \mathbf {x} - \mathbf {y}\Vert \) and \(\Vert \mathbf {x}_{\parallel } - \mathbf {y}_{\parallel } \Vert \le \Vert \mathbf {x} - \mathbf {y}\Vert \), \(\forall \mathbf {x}, \mathbf {y}\).
To verify condition C.1, we need to bound the expected T-step drift for \(Z(\mathbf {X})\). For ease of notation, we denote \({\mathbb {E}}[\ \cdot \ |\mathbf {X}(t) = \mathbf {X}]\) as \({\mathbb {E}}_{\mathbf {X}}[\ \cdot \ ]\).
Since (43) provides an upper bound on the magnitude of the T-step drift, we know that if \(\mathbf {X}(t)\) satisfies \(Z(\mathbf {X}) > D\), then \(Z(\mathbf {X}(t+T))\) is positive, and we may drop the \(\max \{\cdot , 0\}\) expression in the definition of \(Z(\mathbf {X})\) in this case. In other words, for all \(\mathbf {X}\) such that \(Z(\mathbf {X}) > D\), we may write the expected T-step drift as
Therefore, we need only to consider the T-step expected drift of \(\Vert \mathbf {q}_{\perp }\Vert \) and \(\Vert \mathbf {q}_{\parallel }\Vert \).
We first consider the drift of \(\Vert \mathbf {q}_{\perp }\Vert \). The derivation follows along the lines of [15], where the relation in [15, Lemma 4] is used: Let \(V(\mathbf {X}) = ||\mathbf {q}||^2\), \(V_{\parallel }(\mathbf {X}) = ||\mathbf {q}_{\parallel }||^2\), and \(\Delta V\), \(\Delta V_{\parallel }\) denote the one-step drift of V, \(V_{\parallel }\), respectively. Then
The inequality could be derived as follows:
where the inequality follows from the concavity of the square root function: Since \(f(x) = \sqrt{x}, x > 0\) is concave, we have \(f(y) - f(x) \le (y - x) f'(x) = \frac{y-x}{2\sqrt{x}}\). Setting \(x = \Vert \mathbf {q}_{\perp }(t)\Vert ^2\) and \(y = \Vert \mathbf {q}_{\perp }(t+1)\Vert ^2\) gives the inequality. Then, with the orthogonality between \(\mathbf {q}_{\perp }\) and \(\mathbf {q}_{\parallel }\), we have \(\Vert \mathbf {q}_{\perp }\Vert ^2 = \Vert \mathbf {q}\Vert ^2 - \Vert \mathbf {q}_{\parallel }\Vert ^2\), and (45) follows by applying the relation for \(\mathbf {q}_{\perp }(t+1)\) and \(\mathbf {q}_{\perp }(t)\) and then rearranging the terms.
With (45), we have the following inequality for the T-step drift of \(\Vert \mathbf {q}_{\perp }(t)\Vert \):
We now derive bounds for \(\Delta V\) and \(\Delta V_{\parallel }\):
where (a) follows from \({\mathbb {E}}[ a_{ij}^2] = \lambda _{ij}^2 + \sigma _{ij}^2\), \(a_{ij}(t)s_{ij}(t) \ge 0\) for all i, j, and \(\sum _{ij} s_{ij}(t) = 1\) for all t.
Suppose g is the sublinear hysteresis function for Adaptive MaxWeight, then by the sublinearity, there exists a constant \(K_{\theta }\) such that \(g(x) < \frac{\theta }{\alpha } x\) for any \(x > K_{\theta }\), where \(\alpha = \frac{8\Vert \varvec{\nu }\Vert }{\nu _{\min }}\). Hence, by the definition of Adaptive MaxWeight, we have, for any \(\mathbf {X}(\tau )\) such that \(\langle \mathbf {q}(\tau ), \mathbf {s}^*(\tau ) \rangle > K_{\theta }\),
From [15, Claim 2], we have \(\Big \langle \mathbf {q}(\tau ), \varvec{\nu }-\mathbf {s}^*(\tau ) \Big \rangle \le -\nu _{\min } \Vert \mathbf {q}_{\perp }(\tau )\Vert \). Therefore,
For \(\Delta V_{\parallel }\), we have
For (b), we use the following properties of the projection onto cone \(\mathcal {K}\): For \(\mathbf {q} \in {\mathbb {R}}^{n^2}\), \(\langle \mathbf {q}_{\parallel }, \mathbf {q}_{\perp } \rangle = 0\), and \(\mathbf {q}_{\perp } \in \mathcal {K}^{\circ }\). Therefore \(\langle \mathbf {q}_{\parallel }(t), \mathbf {q}_{\perp }(t) \rangle = 0\), and \(\langle \mathbf {q}_{\parallel }(t), \mathbf {q}_{\perp }(t+1) \rangle \le 0\).
Applying (47) and (48) in (46), we obtain
where we have used the fact that \(\Vert \mathbf {q}_{\perp }\Vert \ge \theta \Vert \mathbf {q}_{\parallel }\Vert \) implies \(\theta \Vert \mathbf {q}\Vert \le \theta (\Vert \mathbf {q}_{\parallel }\Vert + \Vert \mathbf {q}_{\perp }\Vert ) \le (1+\theta ) \Vert \mathbf {q}_{\perp }\Vert \), and \(\Vert \mathbf {s}(\tau )\Vert \le \sqrt{n}\) for any schedule \(\mathbf {s}(\tau ) \in \mathcal {S}\).
On the other hand, the drift of \(\Vert \mathbf {q}_{\parallel }\Vert \) can be obtained following (48):
where the last inequality follows from the Cauchy–Schwartz inequality.
Now, applying (49) and (50) in (44), we obtain
From [19, Lemma 1], we know that for any fixed \(T>0\), if \(W^*(t) > g^{-1}\Big ( nT(a_{\max }+1) \Big ) + nT\), then at most one reconfiguration could occur within \([t, t+T]\), which gives \(\sum _{\tau =t}^{t+T-1} \mathbb {1}_{r(\tau ) > 0} \le \Delta _r\).
Select \(T = \frac{8 \sqrt{n} \Delta _r}{\nu _{\min }}\), then set \(D = \frac{3}{2}na_{\max }T = \frac{12 n^{3/2} a_{\max } \Delta _r}{\nu _{\min }}\). Then, \(\forall \mathbf {X}\) such that \(Z(\mathbf {X}) > \kappa = \Big \{ D, na_{\max }T + \frac{4(\Vert \varvec{\lambda }\Vert ^2 + \Vert \varvec{\sigma }\Vert ^2 + n)}{\nu _{\min }} , nK_{\theta },\) \(ng^{-1}\Big ( nT(a_{\max }+1) \Big ) + n^2T \Big \}\), and \(\forall \epsilon \) such that \(0 < \epsilon \le \frac{\nu _{\min }}{16\Vert \varvec{\nu }\Vert }\), we have \({\mathbb {E}}_{\mathbf {X}} \Big [ \Delta ^{\mathrm{T}} Z(\mathbf {X}) \Big ] \le -\frac{(1-\theta )\nu _{\min }}{4} \le -\frac{\nu _{\min }}{8}\).
Hence, by Lemma 1, we have \(\forall \epsilon \) such that \(0 < \epsilon \le \frac{\nu _{\min }}{16\Vert \varvec{\nu }\Vert }\),
Letting \(M_{\theta } = \kappa + \frac{16 D^2}{\nu _{\min }}\), we then have the result. \(\square \)
Proof of Proposition 2
Proof
We will derive the lower bound by bounding the expected queue length sum at each input port, in particular, \({\mathbb {E}}\big [ \sum _{j} \bar{q}_{ij}(t) \big ]\) for each input port i.
The queue length dynamics of the coupled queue \(\phi _i(t)\) are given by
where \(v_i(t)\) is the unused service and satisfies \(\phi _i(t+1)v_i(t) = 0\).
We may show by induction that \({\mathbb {E}}\big [ \sum _{j} \bar{q}_{ij}(t) \big ] \ge {\mathbb {E}}\big [ \bar{\phi }_i(t) \big ]\). It then remains to lower bound \({\mathbb {E}}\big [ \bar{\phi }_i(t) \big ]\). We consider the expected drift of \(\big ( \bar{\phi }_i(t) \big )^2\) as follows:
where the last equality follows from the independence between the queue length process \(\phi _i(t)\) and the schedule reconfiguration decision. We have
Considering the drift of \(\bar{\phi }_i(t)\), we can derive \({\mathbb {E}}\big [ v_i(t) \big ]\) as follows:
We thus have
Using \({\mathbb {E}}\big [ \sum _{j} \bar{q}_{ij}(t) \big ] \ge {\mathbb {E}}\big [ \bar{\phi }_i(t) \big ]\), and summing over each input port, we obtain
\(\square \)
Proof of Proposition 3
Proof
Let \(t_k, k = 1, \ldots \), denote the \(k\mathrm{th}\) schedule reconfiguration time. By the assumption that the system state \(\mathbf {X}(t) = \big ( \mathbf {q}(t), \mathbf {s}(t), r(t) \big )\) converges in distribution to a steady-state random vector \(\bar{\mathbf {X}} = \big ( \bar{\mathbf {q}}, \bar{\mathbf {s}}, \bar{r} \big )\), we may consider a renewal-reward process associated with \(\mathbf {X}(t)\) where \(t_k\) are the arrival epochs, while the reward function is the sum of queue lengths given by \(R(t) = \sum _{ij} q_{ij}(t)\).
From renewal-reward theory, we have
For any (i, j) such that \(S_{ij}(t_k) = 0\), we have \(q_{ij}(t) = q_{ij}(t_k) + \sum _{\tau = t_k}^{t-1} a_{ij}(\tau ) \ge \sum _{\tau = t_k}^{t-1} a_{ij}(\tau )\), while for any (i, j) such that \(S_{ij}(t_k) = 1\), we have \(q_{ij}(t) \ge 0\).
We then have a lower bound on the expected queue length sum as follows:
\(\square \)
Proof of Proposition 4
Proof
Given the period T, we consider the Markov chain \(\mathbf {X}(t)\) being sampled at times \(t_k = kT, k = 0, 1, \ldots \). Since the Fixed Frame MaxWeight policy stabilizes the system if \(T > \frac{\Delta _r}{\epsilon }\), we know that \(\mathbf {X}(t)\) converges to a steady-state distribution, and so does \(\mathbf {X}(t_k)\). Let \(\bar{\mathbf {X}} = (\bar{\mathbf {q}}, \bar{\mathbf {s}}, \bar{r})\) and \(\hat{\mathbf {X}} = (\hat{\mathbf {q}}, \hat{\mathbf {s}}, \hat{r})\) denote the steady-state distribution of \(\mathbf {X}(t)\) and \(\mathbf {X}(t_k)\), respectively. By the assumption on the maximum arrival, we immediately have that \({\mathbb {E}}[ \sum _{ij} \bar{q}_{ij} ] \le {\mathbb {E}}[ \sum _{ij} \hat{q}_{ij} ] + n^2a_{\max }T\). It then remains to bound \({\mathbb {E}}[ \sum _{ij} \hat{q}_{ij} ]\) following the similar procedures in [15]:
-
1.
Derive an upper bound on \({\mathbb {E}}[\Vert \mathbf {q}_{\perp }(t_k)\Vert ^2]\).
-
2.
Derive the queue length upper bound which depends on \({\mathbb {E}}[\Vert \mathbf {q}_{\perp }(t_k)\Vert ^2]\).
Consider the Lyapunov function \(Z(\mathbf {X}) = \Vert \mathbf {q}_{\perp }\Vert \). By the assumption on the maximum arrival, we have
For the expected drift at steady state, we have
where \(\Delta V(\mathbf {X})\) and \(\Delta V_{\parallel }(\mathbf {X})\) are the drift of Lyapunov functions \(V(\mathbf {X}) = \Vert \mathbf {q}\Vert ^2\) and \(V_{\parallel }(\mathbf {X}) = \Vert \mathbf {q}_{\parallel }\Vert ^2\), respectively. Now, for each \(\tau \in [t_k, t_{k+1}-1]\), we have
and
We then have the expected drift of \(Z(\mathbf {X})\) at steady state given by
where the last inequality follows from \(\epsilon \le \frac{\nu _{\min }}{4\Vert \varvec{\nu }\Vert }\) and \(T > \frac{\Delta _r}{\epsilon } \ge \frac{4\Vert \varvec{\nu }\Vert \Delta _r}{\nu _{\min }} \ge \frac{4\sqrt{n}\Delta _r}{\nu _{\min }}\).
Let \(\kappa = \frac{2(\Vert \varvec{\lambda }\Vert ^2 + \Vert \varvec{\sigma }\Vert ^2 + n)}{\nu _{\min }} + nT(\frac{2(a_{\max }+1)}{\nu _{\min }} + 1)\) . We have that \(\Vert \mathbf {q}(t_k)\Vert > \kappa \) implies
then using [15, Lemma 3] with \(D = na_{\max }T\) and \(\eta = \frac{\nu _{\min }T}{4}\), we have
Letting \(M = \frac{ 4na_{\max } + 4(\Vert \varvec{\lambda }\Vert ^2 + \Vert \varvec{\sigma }\Vert ^2 + 2n) + 16\sqrt{2} n^2a_{\max }^2 }{\nu _{\min }} + 2n + 4\sqrt{2} na_{\max }\), we have \({\mathbb {E}}\Big [ \Vert \mathbf {q}_{\perp }(t_k)\Vert ^2 \Big ] \le T^2M^2\).
Consider the Lyapunov function \(W(\mathbf {X}) = \sum _{i} \Big ( \sum _{j} q_{ij} \Big )^2 + \sum _{i} \Big ( \sum _{i} q_{ij} \Big )^2 - \frac{1}{n} \Big ( \sum _{ij} q_{ij} \Big )^2\), and set the corresponding Lyapunov drift at steady state to zero, \({\mathbb {E}}\Big [ W(\mathbf {X}(t_{k+1})) - W(\mathbf {X}(t_{k})) \Big ] = 0\). We then have \(T_1 = T_2 + T_3 + T_4\), where
Hence, we have the upper bound
\(\square \)
Rights and permissions
About this article
Cite this article
Wang, CH., Maguluri, S.T. & Javidi, T. Heavy traffic queue length scaling in switches with reconfiguration delay. Queueing Syst 98, 49–93 (2021). https://doi.org/10.1007/s11134-021-09695-x
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-021-09695-x